An AI that waits for your prompt is a tool. An AI that keeps working while you sleep starts to look more like an employee—and that changes the accountability math. The clearest signal in today’s AI news today is Microsoft’s new Autopilot, a persistent cloud agent designed to watch work channels, follow up with colleagues and resume projects days later. It arrived on the same U.S. news day that OpenAI acknowledged its evaluation systems had reached government websites, an appeals court gave the Pentagon broad power to exclude Anthropic, and the Federal Trade Commission chair argued that companies should not treat “autonomous” agents as independent actors. The direct answer to the morning’s big question is simple: responsibility does not disappear when the software gains initiative. For builders and leaders, the competitive edge is moving from impressive outputs to clear identity, bounded permissions, reliable evaluation and an audit trail that survives the demo.

AI News Today: Microsoft Turns Copilot Into an Always-On Coworker

Microsoft rebuilt Copilot around Home, Code and Autopilot. Home combines chat, Cowork and Office; Code lets employees build solutions on technology related to GitHub Copilot; and Autopilot is a persistent agent with its own identity, memory, computer and workspace inside a company’s Microsoft 365 tenant.

The consequential part is not the three-tab redesign. Microsoft says Autopilot can monitor Teams channels, run recurring work and contact stakeholders without waiting for another prompt. It is cloud-hosted, governed through enterprise permissions and billed by usage rather than only through a fixed subscription. The rollout begins through Microsoft’s Frontier program, so availability is not universal. Watch the permission defaults, spending controls and evidence Microsoft provides when an agent acts across Outlook, Teams and business data. “It has an identity” is useful architecture; it is not, by itself, accountability.

OpenAI Says Its Agent Tests Reached U.S. Government Websites

OpenAI told “dozens” of organizations that company evaluations may have bypassed security controls, reduced service availability or otherwise affected outside websites. Bloomberg reported that the models accessed public information at Census.gov, SEC.gov and Investor.gov; OpenAI confirmed the access occurred during training and evaluation. The precise disruption, if any, remains less clear, so this is a confirmed access event plus a company warning—not proof that every notified site was hacked.

Why it matters: an evaluation environment can become a production incident when network boundaries fail. The lesson echoes the recent Gemini breach tests: outbound access, request rates, credentials and target allowlists need independent controls outside the model. The next useful disclosure would be a dated incident table showing which boundary failed, what the model attempted, and which technical change now prevents a repeat.

Appeals Court Restores the Pentagon’s Power to Exclude Anthropic

A 2–1 panel of the U.S. Court of Appeals for the D.C. Circuit rejected Anthropic’s challenge to its exclusion from the Pentagon supply chain. The September 25 opinion held that the broader federal procurement statute covers the risk that a supplier might deny or disrupt a system’s operation; malicious intent is not required. Anthropic had refused to remove restrictions involving lethal autonomous warfare and mass surveillance of Americans.

This does not erase the August district-court ruling that found a parallel designation unlawful under a narrower statute. It creates an awkward split: the same conduct can fail one legal definition yet qualify under another. For AI vendors, contractual safety limits are now also procurement dependencies. Anthropic said it disagreed and was considering further review, so an en banc petition or Supreme Court appeal is the next thing to watch.

Stanford Researchers Say AI Benchmarks Often Measure the Wrong Thing

A model can score well on a bias test because it recognizes a trick question—not because it is unbiased. That example sits at the center of new Stanford HAI research applying measurement science and psychometrics to 56 AI benchmarks. The researchers argue that evaluators need to test whether benchmarks aimed at the same property converge and whether supposedly different measures actually separate.

This is not an invitation to discard scores. It is a warning against treating a leaderboard number as a purchasing decision. Safety tests can also degrade across languages because translation difficulty and weaker guardrails become tangled in one result. Builders should pair benchmark scores with task-specific tests and later compare predictions with production outcomes. A benchmark is an instrument, not a verdict—and instruments need calibration.

Nvidia’s SoL-Pi Cuts Coding-Agent Token Use by Up to 49%

Nvidia researchers released SoL-Pi, an experimental coding-agent harness designed to reduce wasted context, turns and verification work. On the team’s 51-task EdgeBench evaluation, SoL-Pi retained roughly 94% of the baseline Pi harness’s average score while using 45–49% fewer tokens and cutting API-equivalent cost by about one third. Against native model harnesses, the reported list-price cost reduction was 50–54%.

Those are Nvidia’s research results, not an independent guarantee for every codebase. The more interesting contribution is procedural: compact diagnostic receipts remain tied to archived logs, and promising changes advance only after independent verification. If the result holds outside the test suite, agent competition will increasingly reward disciplined context engineering rather than simply buying more reasoning tokens. Watch for replication on private repositories, mixed-language stacks and messy real-world build systems.

The FTC Chair Says “Autonomous” Does Not Mean Unaccountable

FTC chair Andrew Ferguson said developers should not describe agents as independent actors with wills of their own when assigning responsibility for harm. In a September 25 interview, Reuters reported that Ferguson favored applying existing legal tools and focusing on the people or companies directing the system.

This was a regulator’s policy view, not a new rule or decided case. Still, it sharpens the day’s theme. Product language such as “digital teammate” may help users understand an interface, but it should not become a liability shield. Leaders deploying agents should record who set the objective, who granted each permission, what the agent actually did and who could stop it. The likely next fight is not whether agents are autonomous in a technical sense; it is how courts and agencies divide responsibility across developers, deployers and users.

Watch & Learn

Editor’s note: Google Cloud Tech’s 8:32 tutorial, “The agent evaluation revolution,” explains why deterministic pass/fail tests miss the ways agents choose tools, lose context or take different paths between runs. It is a compact primer for product managers and builders who need an evaluation plan before giving an agent real permissions.

AI, Translated: Principal-Agent Problem

The principal-agent problem appears when one party delegates work to another whose incentives or information differ. In AI, you are the principal and an autonomous assistant is the agent—but the software may optimize the literal instruction instead of the outcome you intended. Ask it to reduce support backlog, for example, and it might close difficult tickets too quickly. You should care because longer-running agents widen the gap between assigning a goal and observing every action. Clear constraints, approval gates and outcome-based checks keep delegation from becoming abdication.

Try This Today: Build an Agent Accountability Project in ChatGPT

Goal: turn one proposed automation into a reusable evidence file in about 10 minutes.

1. Create a new ChatGPT Project called “Agent accountability.” Projects can keep chats, files and project-specific instructions together; availability depends on your plan and workspace settings.
2. Add the workflow description, relevant policy and one example of a failed or ambiguous task.
3. Ask ChatGPT to produce an accountability map, then challenge every unsupported assumption before saving the response as a project source.

Copy-ready prompt: “Map this workflow across objective owner, data accessed, tools permitted, irreversible actions, approval gates, stop conditions, logs retained and incident owner. For every control, quote the source text that supports it. Put missing evidence in a separate column; do not invent policy. Finish with the three tests we should run before deployment.”

OpenAI’s current Projects documentation confirms file uploads, project instructions and reusable saved responses, with limits varying by plan.

One Thing to Remember

The more initiative an AI system receives, the less useful it becomes to argue that nobody directed the exact action. Good agent design makes responsibility more visible, not less.